Understanding AI Agents and Agentic Workflows
Discover how AI agents make decisions and take actions, and learn to create a structured and validated workflow without coding.
What you'll learn
- Distinguish between a chat assistant, a fixed automation, an agentic workflow, and an AI agent based on who determines the subsequent action and the extent of their changes
- Explain how goals, directives, models, context, resources, state, and stopping criteria integrate within an agent's design
- Follow the proposed action of an agent through the selection of tools, permission verifications, execution, outcomes, and subsequent decisions
- Determine the suitable level of autonomy for workplace activities considering clarity, evidence, impact, reversibility, and exceptions
- Design a bounded agentic workflow incorporating inputs, tool authorizations, decision points, human approvals, and exception routes
- Evaluate when a single agent is sufficient or when multiple agents are necessary, and create clear handoff agreements between process steps
- Recognize potential risks from untrusted sources, prompt injections, sensitive data, inaccurate outputs, and excessive access rights
- Implement practical safeguards: least privilege policies, approval checks, evidence validations, retry limits, escalation procedures, and activity documentation
- Prepare a pilot program with relevant test scenarios, acceptance criteria, metrics, fallback plans, and decisions to proceed, revise, or halt
- Compose a one-page agent workflow brief and justify whether a case necessitates an agent, a more straightforward workflow, or human involvement.
Course outline
3 sections · 10 lessons · 1h 38m total
Understand Agents and Their Components
- From Answering Questions to Taking Actions Free Preview On Udemy 09:23
- The Building Blocks of an Agent 09:38
- Follow the Action and Observation Loop 09:04
Design a Useful Workflow
- Choose the Right Task and Level of Autonomy Free Preview On Udemy 09:56
- Map a Bounded Agentic Workflow 10:40
- Coordinate Work Without Losing Control 09:22
Control, Evaluate and Pilot
- Protect Information and Untrusted Content Free Preview On Udemy 08:55
- Bound Actions with Permissions and Approvals 09:31
- Test the Workflow Before Expanding It 10:04
- Final Case: Present Your Agent Workflow Brief 10:59
About this course
This course explores the topic of artificial intelligence.
Someone in your organization has likely already remarked, "We should implement an AI agent for that task." It seems straightforward. However, the term encompasses a range of systems, from a simple tool that drafts responses to one capable of altering data, ordering supplies, and spending money autonomously. Before initiating any project, it's crucial to consider the right questions: Does this function truly require an agent? What actions should it perform autonomously? How will it respond to unforeseen circumstances?
This course will guide you in addressing these queries. You will track a hypothetical field services company as its team transitions from utilizing a chat assistant to developing a regulated request-processing workflow, making design choices step by step. Diagrams, practical examples, and pause-and-decide exercises will be employed, with no coding skills or software installation necessary.
Your learning objectives include:
- Differentiating between a chat assistant, fixed automation, agentic workflow, and AI agent based on their behaviors, not just their names
- Identifying the seven essential components of an agent and understanding why the model proposes actions while the system verifies and executes
- Understanding the action loop: selecting a tool, verifying permissions, executing it, reviewing the result, deciding on the next action, and knowing when to halt
- Selecting the appropriate degree of autonomy for each task, from mere lookups to drafts that require approval to actions within set boundaries
- Constructing a bounded workflow that includes decision points, necessary human approvals, exception pathways, and established completion standards
- Determining whether a single agent suffices or if multiple agents are warranted, and drafting clear handoff contracts for transitions between tasks
- Safeguarding data against untrusted sources and prompt injections, while ensuring sensitive information is secured
- Restricting actions to minimum privileges, incorporating approval gates, evidence verification, retry constraints, and maintaining activity logs
- Conducting tests with representative scenarios, developing acceptance standards, and outlining a small pilot program with a contingency plan
The course culminates in a final project where you will draft a one-page agent workflow brief for a specific situation, then compare it to a model answer and acceptable simplified options. The risk language aligns with contemporary security community recommendations, including the OWASP Top 10 lists pertinent to LLM and agentic applications. You will receive five downloadable resources in Word, Excel, and PDF formats: the case pack, a workflow mapping worksheet, a testing matrix, a template for the agent workflow brief, and a checklist for agent design.
Who this course is for
- Professionals in business and team leaders tasked with evaluating or approving AI agents and AI-driven workflows
- Owners of processes looking to securely automate routine tasks with defined limits, approvals, and exception management
- IT personnel requiring a common, vendor-neutral approach to define agent projects alongside business partners
- Individuals seeking to comprehend how AI agents make decisions and the extent to which human supervision is necessary.